TWI892389B - Method and processing circuit for performing wake-up control on voice-controlled device with aid of detecting voice feature of self-defined word - Google Patents

Abstract

A method for performing wake-up control on a voice-controlled device with aid of detecting voice feature of self-defined word and an associated processing circuit are provided. The method may include: performing feature collection on audio data of at least one audio clip to generate at least one feature list of the at least one audio clip, to establish a feature-list-based database in the voice-controlled device; performing the feature collection on audio data of another audio clip to generate another feature list of the other audio clip; and performing at least one screening operation on at least one feature in the other feature list according to the feature-list-based database to determine whether the other audio clip is invalid, to selectively ignore the other audio clip or execute at least one subsequent operation, where the at least one subsequent operation includes waking up the voice-controlled device.

Description

Method and processing circuit for waking up a voice-controlled device by detecting voice characteristics of custom words

The present invention relates to voice control technology, and more particularly to a method and processing circuit for waking up a voice-controlled device by detecting the voice features of a self-defined word.

According to related technologies, biometric recognition systems can be used to activate user devices to enhance convenience and security, but typically require a remote system with powerful computing capabilities. For example, to accurately identify a speaker, the various conditions involved in the design of an artificial intelligence speaker recognition system can vary with language characteristics, speaking habits, gender and age, vocal structure, etc. Therefore, building a speaker model requires a large amount of appropriate voice data for training the artificial intelligence speaker recognition system to successfully perform automatic recognition. Since the remote system, such as the artificial intelligence speaker recognition system, is usually connected to a wired or wireless network, its availability can be affected by network interruptions. Therefore, a novel method and related architecture are needed to achieve startup control of any remote system without relying on powerful computing capabilities, with no or minimal side effects.

One of the objectives of the present invention is to provide a method and processing circuit for waking up a voice-controlled device by detecting the voice characteristics of custom words, thereby resolving problems in related technologies and preventing thieves or young children from accidentally activating the voice-controlled device.

The present invention provides a method for performing wake-up control on a voice-controlled device by detecting voice features of custom words, wherein the method may include: in a registration phase among multiple phases, performing feature collection on audio data of at least one audio clip to generate at least one feature list of the at least one audio clip, and establishing a feature-list-based database in the voice-controlled device, wherein the at least one audio clip carries at least one custom word, the feature-list-based database includes the at least one feature list, any one of the at least one feature list includes multiple features of an audio clip corresponding to one of the at least one audio clip, and the multiple features respectively belong to a plurality of predetermined types of features. features); in an identification phase among the multiple phases, performing the feature collection on audio data of another audio segment to generate another feature list for the other audio segment; and in the identification phase, performing at least one filtering operation on at least one feature in the other feature list based on the database based on the feature list to determine whether the other audio segment is invalid, thereby selectively ignoring the other audio segment or performing at least one subsequent operation, wherein the at least one subsequent operation includes waking up the voice control device.

The present invention also provides a processing circuit for performing wake-up control on a voice-controlled device by detecting the voice characteristics of custom words.

Among the many advantages of the present invention is that the method and processing circuit of the present invention can determine whether an unknown speaker is invalid based on the feature list-based database, selectively ignoring the audio clip or waking/activating the voice-controlled device, without requiring a connection to any remote system to obtain any voice data for such determination. For example, the method does not require determining the characters in a custom word, thus eliminating the need to connect to any cloud database via any network to obtain large amounts of voice data. Furthermore, the method and processing circuit of the present invention can implement a compact, fast, secure, and reliable voice-controlled processing system with minimal or no side effects.

100: Voice-controlled device

110: Processing circuit

111: Processing Module

112: Threshold initialization processing module

113: Short-Term Energy (STE) and Zero Crossing Rate (ZCR) Processing Module

114: Voice type classification processing module

115: Pitch Processing Module

116: Feature List Processing Module

120: Audio input device

130: Audio data conversion interface circuit

140: Storage device

141: Code

142: Feature List-Based Database

610,620: Audio analysis operation

711: Opening audio segment

719: Ending audio segment

720: Main audio segment

f'1~f'18,f1~f16: audio frames

Seg1~Seg9: audio segments

y1,y2: audio sample

S10~S13, S20~S23, S30~S36, S41~S45, S50~S52: Steps

STE( ): Short-term energy

STE_th: Short-term energy threshold

X ₁ ,X ₂ : axis

ZCR( ):Zero Crossing Rate

ZCR_th: Zero crossing rate threshold

Figure 1 is a schematic diagram of a voice control device according to an embodiment of the present invention.

Figure 2 illustrates a workflow of a registration and recognition control scheme for a method of waking up a voice-controlled device by detecting the voice characteristics of custom words according to an embodiment of the present invention.

FIG3 illustrates the audio and related signals involved in a feature collection control scheme of the method according to an embodiment of the present invention.

Figure 4 illustrates an audio pre-processing operation involved in the threshold initialization process in the feature collection control scheme according to an embodiment of the present invention.

Figure 5 illustrates the audio samples and audio frames involved in the threshold initialization process in the feature collection control scheme according to an embodiment of the present invention.

Figure 6 illustrates the audio analysis operation in the feature collection control scheme according to one embodiment of the present invention.

Figure 7 illustrates the audio frames and audio segments involved in the speech type classification in the feature collection control scheme according to an embodiment of the present invention.

Figure 8 illustrates a workflow of a speaker identification control scheme of the method according to an embodiment of the present invention.

FIG9 illustrates the data points involved in the k-nearest neighbors (k-NN) algorithm classifier in the speaker identification control scheme according to one embodiment of the present invention.

Figure 10 shows a flow chart of the method according to one embodiment of the present invention.

FIG1 is a schematic diagram of a voice-controlled device 100 according to an embodiment of the present invention. Voice-controlled device 100 may include a processing circuit 110 (e.g., a voice-controlled processing circuit), an audio input device 120, an audio data conversion interface circuit 130, and at least one storage device 140. Processing circuit 110 may include multiple processing modules 111 for performing operations of processing circuit 110. These multiple processing modules 111 may include a threshold initialization processing module 112, a short-term energy (STE) and zero-crossing rate (ZCR) processing module 113, a voice type classification processing module 114, a pitch processing module 115, and a feature list processing module 116. The feature list processing module 116 can perform feature-list-related processing, while at least one other processing module, such as the threshold initialization processing module 112, the short-term energy and zero-crossing rate processing module 113, the speech type classification processing module 114, and the pitch processing module 115, can be used to collect features.

For example, processing circuit 110 may be implemented by a processor or microprocessor, audio input device 120 may be implemented by a microphone or headset, audio data conversion interface circuit 130 may be implemented by an amplifier or analog-to-digital converter, and storage device 140 may be implemented by an electrically erasable programmable read-only memory (EEPROM) or flash memory. Examples of voice-controlled device 100 include, but are not limited to, voice-controlled locks such as door locks and car locks, as well as voice-controlled toys. The multiple processing modules 111 may represent multiple program modules running on the processing circuit 110, wherein the voice-controlled device 100 may load the program code 141 onto the processing circuit 110 to form the multiple program modules. In some embodiments, the multiple processing modules 111 may represent multiple sub-circuits of the processing circuit 110.

FIG. 2 illustrates a workflow of a registration and recognition control scheme for a method of performing wakeup control on a voice-controlled device (e.g., voice-controlled device 100) by detecting voice features of custom words according to an embodiment of the present invention. The processing circuit 110 may execute a registration procedure for speakers A and B respectively in a registration phase among the multiple phases, and may execute an identification procedure for speaker U (e.g., an unknown speaker) in an identification phase among the multiple phases, wherein the registration procedure for speaker A may include steps S10 to S13, the registration procedure for speaker B may include steps S20 to S23, and the identification procedure for speaker U may include steps S30 to S34 and step S35 or S36. In the registration phase, the processing circuit 110 may record a plurality of audio clips {Audio_Clip} (e.g., audio clip {Audio_Clip _A } of speaker A and audio clip {Audio_Clip _B } of speaker B), and perform feature collection on the respective audio data {Audio_Data} of the audio clips {Audio_Clip} (e.g., audio data {Audio_Data _A } of audio clip {Audio_Clip _A } and audio data {Audio_Data _B } of audio clip {Audio_Clip _B }) to generate respective feature lists {L} of the audio clips {Audio_Clip} (e.g., feature lists {L A} of audio clip {Audio_Clip _A } and {L _A } of audio clip {Audio_Clip _B}) . } ₎ to establish a feature list-based database 142 in the voice control device 100. Feature list-based database 142 may include feature lists {L}. Any feature list L in feature lists {L} may include multiple features of an audio clip Audio_Clip corresponding to one of the audio clips {Audio_Clip}, and these multiple features may belong to multiple predetermined types of features. For example, the audio clip {Audio_Clip} may contain at least one custom word W. In particular, each audio clip Audio_Clip _A in the audio clips {Audio_Clip _A } may contain a custom word W _A , and each audio clip Audio_Clip _B in the audio clips {Audio_Clip _B } may contain a custom word W _B. In addition, in the recognition phase, the processing circuit 110 may record another audio clip Audio_Clip (e.g., audio clip Audio_Clip _U of speaker U), and perform the feature collection on the audio data Audio_Data of the another audio clip Audio_Clip (e.g., audio data Audio_Data _U of audio clip Audio_Clip _U ) to generate another feature list L of the another audio clip Audio_Clip (e.g., feature list L _U of audio clip Audio_Clip _U) . ), and performing at least one filtering operation on at least one feature in the other feature list L based on the feature list-based database 142 to determine whether the other audio clip Audio_Clip is invalid, so as to selectively ignore the other audio clip Audio_Clip or perform at least one subsequent operation without having to connect to any cloud database through any network to obtain any voice data for determining which words are included in the above-mentioned at least one custom word, wherein the above-mentioned at least one subsequent operation may include waking up/activating the voice-controlled device 100.

In step S10, the processing circuit 110 may begin to execute the registration process for speaker A.

In step S11 , the processing circuit 110 may record a corresponding audio clip Audio_Clip _A in the audio clips {Audio_Clip _A } to record the self-defined words W _A , such as Speaker-A-defined words W _A , as a wake-up word dedicated to Speaker A.

In step S12, the processing circuit 110 may perform feature collection on the audio data Audio_Data _A corresponding to the audio clip Audio_Clip _A to obtain a plurality of features of the corresponding audio clip Audio_Clip _A. The processing circuit 110 may re-enter step S11 as indicated by the dashed arrow to repeatedly execute steps S11 and S12 to respectively perform feature collection on the audio data {Audio_Data _A } of the audio clip {Audio_Clip _A } to obtain respective features of the audio clip {Audio_Clip _A }. For example, processing circuit 110 may provide a user interface such as a record button and a stop button. Speaker A can press the record button to record the custom word W _A at the same pitch and volume, then press the stop button. Processing circuit 110 may detect certain voice characteristics of the corresponding audio clip Audio_Clip _A. If these voice characteristics meet predetermined recording rules, processing circuit 110 may record the corresponding audio clip Audio_Clip _A. Otherwise, processing circuit 110 may notify speaker A to record again.

In step S13, the processing circuit 110 may generate a feature list _LA according to the features of the corresponding audio clip Audio_Clip _A , and in particular, generate a feature list { _LA } for the audio clip {Audio_Clip _A } according to the features of the audio clip {Audio_Clip _A }.

In step S20, the processing circuit 110 may begin to execute the registration process for speaker B.

In step S21 , the processing circuit 110 may record a corresponding audio clip Audio_Clip _B in the audio clip {Audio_Clip B } to record the self-defined words W _B , such as Speaker-B-defined words W _B , as _a wake-up word dedicated to Speaker B.

In step S22, processing circuit 110 may perform feature collection on the audio data Audio_Data _B corresponding to audio clip Audio_Clip _B to obtain multiple features of audio clip Audio_Clip _B. Processing circuit 110 may re-enter step S21, as indicated by the dashed arrow, to repeatedly execute steps S21 and S22 to perform feature collection on the audio data {Audio_Data _B } of audio clip {Audio_Clip _B } to obtain the respective features of audio clip {Audio_Clip _B }. For example, speaker B may press the record button and record the custom word W _B using the same pitch and volume, then press the stop button. The processing circuit 110 may detect certain voice characteristics of the corresponding audio clip Audio_Clip _B. If these voice characteristics meet the predetermined recording rules, the processing circuit 110 may record the characteristics of the corresponding audio clip Audio_Clip _B ; otherwise, the processing circuit 110 may notify speaker B to record again.

In step S23, the processing circuit 110 may generate a feature list L _B according to the features of the corresponding audio clip Audio_Clip _B , and in particular, generate a feature list {L _B } for the audio clip {Audio_Clip _B } according to the features of the audio clip {Audio_Clip _B }.

In step S30, the processing circuit 110 may begin executing the recognition process for speaker U.

In step S31 , the processing circuit 110 may record the audio clip Audio_Clip _U to record any custom word W _U of the speaker U (if the custom word W _U exists).

In step S32 , the processing circuit 110 may perform the feature collection on the audio data Audio_Data _U of the audio clip Audio_Clip _U to obtain a plurality of features of the audio clip Audio_Clip _U.

In step S33, the processing circuit 110 may generate a feature list L _U according to the multiple features of the audio clip Audio_Clip _U.

In step S34, processing circuit 110 may perform speaker identification based on feature list database 142. Specifically, processing circuit 110 may quickly filter one or more features in feature list L _U based on feature list database 142 to determine whether audio clip Audio_Clip _U is invalid. For example, if audio clip Audio_Clip _U is determined to be invalid, this indicates that speaker U is an invalid speaker, and processing circuit 110 may execute step S36. If audio clip Audio_Clip _U is determined to be not invalid, this indicates that speaker U is speaker A or speaker B, and processing circuit 110 may execute step S35.

In step S35, the processing circuit 110 may perform at least one of the subsequent operations described above as an action ( ), in particular, waking up/activating the voice-controlled device 100.

In step S36, the processing circuit 110 may ignore the audio clip Audio_Clip _U.

Because registered speakers can have different custom words {W}, for any registered speaker, as long as the custom word W is not overheard by others, there is a first level of security. Even if the custom word W is overheard by others, there is a second level of security because the voice characteristics will not be able to wake up/activate voice-controlled device 100. If the speaker speaks custom word W with a different tone or volume, the processing circuit 110 will determine that the audio clip Audio_Clip is invalid/unqualified. Therefore, there is no need to worry about daily conversations being recorded and used to forge voice to activate voice-controlled device 100. Furthermore, processing circuit 110 can quickly determine whether speaker U (e.g., an unknown speaker) is invalid based on database 142 based on a feature list, thereby selectively ignoring its audio clip Audio_Clip _U or waking/activating voice-controlled device 100 without requiring a connection to any remote system to obtain any voice data for relevant judgments. Since there is no need to determine the characters in a custom word, processing circuit 110 does not need to connect to any cloud database via any network to obtain large amounts of voice data. Therefore, the method and processing circuit 110 of the present invention can implement a compact, fast, safe, and reliable voice-controlled processing system with no or minimal side effects.

According to some embodiments, one or more steps may be added, deleted, or modified in the workflow shown in FIG. 2 . For example, by performing machine learning, the processing circuit 110 may establish a predetermined classifier corresponding to at least one predetermined model in the voice-controlled device 100 . The dimension of a predetermined space (e.g., a predetermined space expanded by multiple axes {X ₁ , X ₂ , ...}) of the at least one predetermined model may be equal to the feature-type count of the multiple predetermined types of features. Furthermore, the at least one feature in the other feature list L may be at least one of all features in the other feature list L, wherein all features in the other feature list L belong to the multiple predetermined types of features. During the execution of at least one of the aforementioned subsequent operations, the processing circuit 110 may utilize the predetermined classifier to perform machine-learning-based classification based on all of the aforementioned features in the other feature list L to determine whether the speaker U of the other audio clip Audio_Clip is speaker A (e.g., user A) or speaker B (e.g., user B), so as to selectively execute at least one action Action(A) corresponding to speaker A (e.g., user A) or at least one action Action(B) corresponding to speaker B (e.g., user B) in step S35.

FIG. 3 illustrates audio and related signals involved in a feature collection control scheme according to an embodiment of the present invention. The horizontal axis may represent time, measured in milliseconds (ms), and the vertical axis may represent the intensity of audio sample y. According to certain embodiments, audio and related signals, such as short-term energy (STE), zero-crossing rate (ZCR), and classification signals, may be varied. As shown in FIG. 3 , processing circuit 110 may generate a classification signal, indicating whether any of multiple portions of the audio is unvoiced, voiced, or breathy, based on whether the short-term energy (STE) reaches a short-term energy threshold (STE_th) and whether the zero-crossing rate (ZCR) reaches a zero-crossing rate threshold (ZCR_th). For example, the processing circuit 110 may detect that the pitch of the audio signal is equal to 181.07 Hertz (Hz).

FIG. 4 illustrates an audio pre-processing operation involved in the threshold initialization process in the feature collection control scheme according to an embodiment of the present invention, wherein the threshold initialization processing module 112 shown in FIG. 1 may be used to perform the threshold initialization process. The threshold initialization processing module 112 may perform the audio pre-processing operation on an original version of audio sample y (e.g., audio sample y1) to generate a new version of audio sample y (e.g., audio sample y2). Assuming that the sampling frequency of audio sample y1 is 48,000 Hz and the duration (or time length) t _NOISE of the noise portion at the beginning of audio sample y1 is 0.4 seconds (s), the number of samples n is equal to (48,000 * 0.4) = 19,200. According to some embodiments, relevant parameters such as the sampling frequency and duration _tNOISE may be varied. The threshold initialization processing module 112 may calculate the average value MEAN as follows: MEAN = (SUM(y1[0:n-1])/n) = (SUM(y1[0:19199])/19200), where "SUM()" may represent summation. The average value MEAN may indicate signal offset caused by components in the audio input path (e.g., the audio input device 120 and/or the audio data conversion interface circuit 130). The threshold initialization processing module 112 can calibrate the original zero level according to the average value MEAN. Specifically, the average value MEAN is subtracted from all recorded audio samples y1 to generate audio sample y2 as follows: y2[0:N-1]=y1[0:N-1]-MEAN; where "N" can represent the number of samples of audio sample y1 (or y2).

FIG5 illustrates the audio sample {y2} (e.g., audio sample y2[0:n-1]) and audio frame {f'} (e.g., audio frame {f'1, f'2, f'3, ..., f'18}) involved in the threshold initialization processing in the feature collection control scheme according to an embodiment of the present invention, wherein the threshold initialization processing module 112 can subtract the average value MEAN from the original zero level shown in FIG4 to obtain a new zero level. Assuming that the frame size p1 of the noise portion audio frame {f'} (or noise frame {f'}) is equal to 1024 audio samples, the number of frames in the noise portion duration _tNOISE is equal to (n/p1) = (19200/1024) = 18.75~=18, which means that there are at least 18 frames in the duration _tNOISE . According to some embodiments, relevant parameters such as duration _tNOISE , number of samples n, and frame size p1 can be varied. Assuming frame size p = p1, any audio frame f'(i) in audio frame {f'} can contain audio samples {y2[(i-1)*p],...,y2[(i*p)-1]|i=1,...,18}. The threshold initialization processing module 112 may calculate the short-term energy value STE(f'(i)) of the audio frame f'(i) as follows: STE(f'(i))=SUM({y2[x] ² |x=((i-1)*p),...,((i*p)-1)}); where "y2[x] ² " may represent the energy of the audio sample y2[x]. The threshold initialization processing module 112 may calculate the short-term energy threshold STE_th based on the short-term energy value {STE(f'(i))} of the audio frame {f'(i)} as follows: STE_th=MAX({STE(f'(i))})*FACTOR_STE; where "MAX( )" may represent a maximum value, and "FACTOR_STE" may represent a predetermined short-term energy factor. For example, FACTOR_STE=10 is used to determine the short-term energy threshold STE_th for determining whether the speaker is speaking or only has noise. According to some embodiments, the short-term energy threshold STE_th and/or the predetermined short-term energy factor FACTOR_STE can be varied.

In addition, the threshold initialization processing module 112 can calculate the zero crossing rate threshold ZCR_th. Let y3[x] be a function of y2[x] to indicate whether y2[x] is greater than zero as follows: y3[x]=1 if y2[x]>0; and y3[x]=-1 if y2[x] 0.

According to some embodiments, y3[x] and/or related judgment conditions (e.g., y2[x]>0 and/or y2[x] 0) can be varied. The threshold initialization processing module 112 can calculate the zero-crossing rate value ZCR(f'(i)) of the audio frame f'(i) as follows: ZCR(f'(i))=SUM({(|y3[x+1]-y3[x]|)|x=((i-1)*p),…,((i*p)-1)})/2p; where "|y3[x+1]-y3[x]|" can represent the absolute value of (y3[x+1]-y3[x]). According to the above-mentioned predetermined recording rules, the zero-crossing rate of the noise sequence is expected to be large enough, in particular, to reach a predetermined noise sequence zero-crossing rate threshold Noise_Sequence_ZCR_th, which indicates that the noise sequence is qualified as noise for correct threshold initialization processing. In the registration process, the threshold initialization processing module 112 can determine whether to notify the speaker/user being registered (e.g., speaker/user A or speaker/user B) to re-record based on the zero-crossing rate value {ZCR(f'(i))} of each audio frame {f'(i)} and the predetermined noise sequence zero-crossing rate threshold Noise_Sequence_ZCR_th, and the related operations may include: (1) if MIN({ZCR(f'(i))})<Noise_Sequence_ZCR_th, the processing circuit 110 (or the threshold initialization processing module 112) can control the voice control device 100 to notify the speaker/user to re-record; and (2) if MIN({ZCR(f'(i))}) If the predetermined noise sequence zero-crossing rate threshold Noise_Sequence_ZCR_th is equal to 0.3, the threshold initialization processing module 112 may set ZCR_th = MIN({ZCR(f'(i))}), where "MIN()" may represent a minimum value. For example, the predetermined noise sequence zero-crossing rate threshold Noise_Sequence_ZCR_th may be equal to 0.3. In some examples, the predetermined noise sequence zero-crossing rate threshold Noise_Sequence_ZCR_th may be variable.

FIG6 illustrates the audio analysis operations 610 and 620 in the feature collection control scheme according to an embodiment of the present invention, wherein the threshold initialization processing module 112 and the short-term energy and zero-crossing rate processing module 113 shown in FIG1 can be used to perform the audio analysis operations 610 and 620, respectively. When recording the corresponding audio clip Audio_Clip (e.g., the corresponding audio clip Audio_Clip _A or the corresponding audio clip Audio_Clip _B ) to obtain the corresponding audio data Audio_Data of the corresponding audio clip Audio_Clip (e.g., the corresponding audio data Audio_Data _A or the corresponding audio data Audio_Data B), the audio data Audio_Data A and the corresponding audio data Audio_Data _B are recorded. ), the threshold initialization processing module 112 may analyze the first audio data Audio_Data1 of a first partial audio clip Audio_Clip1 (e.g., the noise portion) of the corresponding audio clip Audio_Clip to determine the short-term energy threshold STE_th and the zero-crossing rate threshold ZCR_th based on multiple first audio frames (e.g., audio frame {f'}) of the first audio data Audio_Data1 for further processing of the remaining audio data Audio_Data2 of a remaining partial audio clip Audio_Clip2 of the corresponding audio clip Audio_Clip. As shown in the leftmost portion of FIG. 6 , the beginning of the audio (e.g., audio frame {f'(i)} such as audio frame {f'1, f'2, f'3, ..., f'18}) can be expected to be noise/interference.

In addition, the short-term energy and zero-crossing rate processing module 113 can analyze the remaining audio data Audio_Data2 of the remaining partial audio clip Audio_Clip2 to calculate the short-term energy value {STE()} and zero-crossing rate {ZCR()} of each of the plurality of second audio frames (eg, audio frame {f}) of the remaining audio data Audio_Data2. Based on whether the short-term energy value STE() of any second audio frame (e.g., audio frame f) among the plurality of second audio frames reaches a short-term energy threshold STE_th and whether the zero-crossing rate ZCR() of any second audio frame reaches a zero-crossing rate threshold ZCR_th, the processing circuit 110 (or the voice type classification processing module 114) can determine whether the voice type of any second audio frame is one of a plurality of predetermined voice types (e.g., a non-voice type, a voice type, and a breathy voice type), so as to determine the plurality of features of the corresponding audio clip Audio_Clip based on the voice types of each of the plurality of second audio frames (e.g., audio frame {f}).

Assuming that the frame size p=p2, the short-term energy and zero-crossing rate processing module 113 can calculate the short-term energy value STE(f(j)) and the zero-crossing rate ZCR(f(j)) of any audio frame f(j) in the audio frame {f(j)} (for example, the audio frames {f1, f2, f3, f4, f5, f6, f7, f8, f9, f10, f11, f12, f13, f14, f15, f16}), and based on a set of predetermined classification rules, the processing circuit 110 (or The speech type classification processing module 114 may classify any audio frame f(j) into one of the plurality of predetermined speech types based on the short-term energy value STE(f(j)) and the zero-crossing rate ZCR(f(j)). For example: (1) if STE(f(j))<STE_th, the processing circuit 110 (or the speech type classification processing module 114) may determine that the speech type of any audio frame f(j) is the silent type; (2) if STE(f(j)) STE_th and ZCR(f(j))<ZCR_th, then the processing circuit 110 (or the voice type classification processing module 114) can determine that the voice type of any audio frame f(j) is the voice type; and (3) if STE(f(j)) STE_th and ZCR(f(j)) ZCR_th, the processing circuit 110 (or the speech type classification processing module 114) can determine that the speech type of any audio frame f(j) is the breathy type.

According to some embodiments, the set of predetermined classification rules and related calculations and/or related parameters such as frame size p, the number of audio frames {f(j)}, etc. can be varied.

FIG7 illustrates the audio frames {f(j)} and audio segments {Seg(k)} involved in voice type classification in the feature collection control scheme according to one embodiment of the present invention, wherein the voice type classification processing module 114 shown in FIG1 may be used to perform the voice type classification. Based on the respective voice types of the plurality of second audio frames (e.g., audio frames {f(j)}, such as audio frames {f1, f2, ..., f16}), the voice type classification processing module 114 may classify the corresponding audio segment Audio_Clip into a plurality of audio segments {Seg(k)}, such as audio segments {Seg1, Seg2, Seg3, Seg4, Seg5, Seg6, Seg7, Seg8, Seg9}. Any two adjacent audio frames having the same predetermined audio type in all audio frames of the corresponding audio data Audio_Data may belong to the same audio segment Seg(k). All the audio frames of the corresponding audio data Audio_Data may include the plurality of first audio frames (e.g., audio frame {f'(i)}) and the plurality of second audio frames (e.g., audio frame {f(j)}). A leading audio segment 711 (e.g., audio segment Seg1) in the plurality of audio segments {Seg(k)} may include at least the plurality of first audio frames and may correspond to a first predetermined audio type, such as the silent audio type.

In particular, the speech type classification processing module 114 may calculate the total duration, e.g., duration( ), of at least one primary audio segment 720 (e.g., audio segments {Seg2, Seg3, ..., Seg8}) from the plurality of audio segments {Seg(k)} as one of the plurality of features of the corresponding audio clip Audio_Clip. The at least one primary audio segment 720 may include audio segments from the plurality of audio segments {Seg(k)}, such as audio segments {Seg2, Seg3, ..., Seg8}, excluding the leading audio segment 711 and any trailing audio segments 719 (e.g., audio segment Seg9) corresponding to the first predetermined speech type (e.g., the non-speech type). In addition, the speech type classification processing module 114 can use one or more other processing modules in the processing module 111 to calculate at least one segment-level parameter of each audio segment Seg(k) (for example, each of the audio segments {Seg2, Seg4, Seg6, Seg8}) corresponding to a second predetermined speech type (for example, the speech type) in the multiple audio segments {Seg(k)}, so as to determine at least one parameter of the corresponding audio clip Audio_Clip based on the above at least one segment-level parameter as at least one other feature of the multiple features of the corresponding audio clip Audio_Clip, wherein each of the above audio segments Seg(k) can represent a speech segment Seg(k). For example, the at least one segment-level parameter may include the pitch ( Pitch ), short-term energy ( STE ), and zero-crossing rate ( ZCR ) of each audio segment Seg(k), and the at least one parameter may include the pitch ( Pitch ), short-term energy ( STE ), and zero-crossing rate ( ZCR ) of the corresponding audio clip Audio_Clip. According to certain embodiments, the at least one segment-level parameter and/or the at least one parameter may be variable. For example, the at least one parameter may further include the start time point Pos ( ) of a main audio segment in the at least one main audio segment 720 of the corresponding audio clip Audio_Clip.

The calculation of the pitch Pitch( ) for each of the above-mentioned audio segments Seg(k) (e.g., each of the audio segments {Seg2, Seg4, Seg6, Seg8}) can be described as follows. The speech type classification processing module 114 can use the pitch processing module 115 to calculate the pitch Pitch( ) of the speech segment Seg(k) according to any one of one or more predetermined pitch calculation functions. For example, the one or more predetermined pitch calculation functions can include a first predetermined pitch calculation function such as an auto-correlation function (ACF) ACF( ) and a second predetermined pitch calculation function such as an average magnitude difference function (AMDF) AMDF( ). Assuming that the length of a speech segment Seg(k) is Q (or Q audio samples), the autocorrelation function ACF( ) can be expressed as follows: ACF(m) = Σ _{q = m} ^Q-1 (y2(q) * y2(qm)); or ACF(m) = Σ _{q = m, ..., Q-1} (y2(q) * y2(qm)); where "Σ" represents summation, and "q" represents the sample index of the associated audio sample y2(q), which can be an integer in the interval [m, Q-1]. The pitch processing module 115 can find the maximum ACF(m) in the range _M1 < m < _M2 -1. For example, _M1 may be the number of autocorrelation points to be calculated, corresponding to the minimum pitch period to be searched by the pitch processing module 115, and _M2 may be the number of autocorrelation points to be calculated, corresponding to the maximum pitch period to be searched by the pitch processing module 115. Furthermore, the mean amplitude difference function (AMDF) can be expressed as follows: AMDF(m) = Σq ₌₀ ^Q-1 (y2(q)-y2(q+m)); or AMDF(m) = _{Σq=0,...,Q-1} (y2(q)-y2(q+m)).

Wherein "Σ" may represent summation, and "q" may represent the sample index of the relevant audio sample y2(q), and may be an integer in the interval [0, Q-1]. The pitch processing module 115 may find the minimum AMDF(m) in the range of _M1 <m< _M2-1 .

A speech segment Seg(k) (whose length is equal to Q) can be considered a periodic signal with a period T ₀ = m, and a fundamental frequency f ₀ equal to (1/T ₀ ). For example, when the sampling frequency fs is equal to 48,000 Hz, the pitch processing module 115 can calculate the period T ₀ and fundamental frequency f ₀ as follows: T ₀ = m = 250 (samples) = (250/48,000) (seconds); and f ₀ = (1/T ₀ ) = (fs/m) = (48,000/250) = 192 (Hz).

According to certain embodiments, the one or more predetermined pitch calculation functions and/or related parameters may be varied. Generally speaking, the pitch (Pitch) may range from 60 to 300 Hz, with the pitch (Pitch) for men ranging from 60 to 180 Hz and the pitch (Pitch) for women ranging from 160 to 300 Hz. For example, when the sampling frequency fs is 48,000 Hz: if _f0 = 60 Hz, then _T0 = (48,000/60) = 800; and if _f0 = 300 Hz, then _T0 = (48,000/300) = 160; thus, the pitch period may be 160 to 800 samples. Therefore, the pitch processing module 115 may set _M1 = 160 and _M2 = 800.

The calculation of the short-term energy value STE( ) and the zero-crossing rate ZCR( ) for each of the above audio segments Seg(k) (for example, there is a speech segment Seg(k), such as each of the audio segments {Seg2, Seg4, Seg6, Seg8}) can be explained as follows. The speech type classification processing module 114 can use the short-term energy and zero-crossing rate processing module 113 to calculate the short-term energy value STE(Seg(k)) and the zero-crossing rate ZCR(Seg(k)) of the speech segment Seg(k) as follows: STE(Seg(k))=AVG({STEf(j))|j=j1(k),...,j2(k)}); ZCR(Seg(k))=AVG({ZCRf(j))|j=j1(k),...,j2(k)}); where "AVG( )" may represent the average, "j1(k)" may represent the index of the beginning audio frame f(j1(k)) of the speech segment Seg(k), and "j2(k)" may represent the index of the ending audio frame f(j2(k)) of the speech segment Seg(k).

According to some embodiments, the processing circuit 110 (or the feature list processing module 116) may generate a feature list {L}, such as feature lists { _LA } and { _LB }, according to a predetermined feature list format. For example, the plurality of predetermined features may include short-term energy value STE(), zero-crossing rate ZCR(), start time point Pos(), pitch Pitch(), and duration Duration(), and the predetermined feature list format may represent a feature list format (STE(), ZCR(), Pos(), Pitch(), Duration()) containing short-term energy value STE(), zero-crossing rate ZCR(), start time point Pos(), pitch Pitch(), and duration Duration(). According to some examples, the plurality of predetermined types of features and/or the format of the predetermined feature list may be varied.

Table 1 shows examples of temporary feature lists {L_tmp _A } and {L_tmp _B } for speakers A and B, while Table 2 shows examples of feature lists { _LA } and { _LB } for speakers A and B. Processing circuit 110 (or short-term energy and zero-crossing rate processing module 113) can find the maximum value of the short-term energy values {STE(Seg(k))} of each speech segment {Seg(k)} (e.g., audio segments {Seg2, Seg4, Seg6, Seg8}) as the short-term energy value STE( ) of the corresponding audio clip Audio_Clip. In particular, the processing circuit 110 (or the feature list processing module 116) can record the short-term energy value STE(Seg(k)), the zero-crossing rate ZCR(Seg(k)), the starting time point Pos(Seg(k)), and the pitch Pitch(Seg(k)) of the speech segment Seg(k) having the maximum value as the short-term energy value STE( ), the zero-crossing rate ZCR( ), the starting time point Pos( ) and the pitch Pitch( ) of the corresponding audio clip Audio_Clip, and record the duration Duration( ) of the audio clip Audio_Clip to establish a temporary feature list {L_tmp} (for example, temporary feature lists {L_tmp _A } and {L_tmp _B }), and normalizes the temporary feature list {L_tmp} (e.g., temporary feature list {L_tmp _A } and {L_tmp _B }) to generate a feature list {L} (e.g., feature list {L _A } and {L _B }).

For example, the feature list processing module 116 may perform the feature list related processing to generate the temporary feature lists {L_tmp _A } and {L_tmp _B } as shown in Table 1: {{(0.0001776469,0.057857143,2800,111.7718733,39200),(0.0000499814,0.021830357,25200,109.6723773,39200),...,(0.0000897361,0.059107143,5600,112.5243115,44800)}, {(0.000117627,0.044642857,2800,189.5970772,42000),(0.0003191778,0.036785714,44800,182.5511925,44800),...,(0.0001378857,0.033214286,44800,178.4916067,44800)}}, and convert the temporary feature lists {L_tmp _A } and {L_tmp _B } into feature lists {L _A } and {L _{B }} respectively. }As shown in Table 2, the feature list {{(0.0823451200,1.45292435,-0.95713666,-0.97132086,-1.3764944),(-1.4695843500,-1.75679355,0.27787838,-1.02837763,-1.37649 44),...,(-0.9863176100,1.56429003,-0.80275978,-0.95087229,0.91766294)},{(-0.6472588500,0.27563005,-0.95713666,1.14368901,-0.22941573), (1.8028253900,-0.42438277,1.35851655,0.95220714,0.91766294),...,(-0.4010006400,-0.74257042,1.35851655,0.84188216,0.91766294)}}. In other examples, the temporary feature lists {L_tmp _A } and {L_tmp _B } and the feature lists {L _A } and {L _B } may be varied.

In either Table 1 or Table 2, each row can be considered as a sample, and each column except row 0 (used to mark language A or B) can be considered as a feature F(Col). The feature list processing module 116 can perform the above-mentioned normalization on rows 1 to 5 of Table 1 (e.g., features {F1(1), F1(2), F1(3), F1(4), F1(5)}) to generate rows 1 to 5 of Table 2 (e.g., features {F2(1), F2(2), F2(3), F2(4), F2(5)}). For the transformed feature F2 in Table 2, the mean is 0 and the standard deviation is 1, which can reduce the influence of outliers. For example, the related operations may include: (1) for each feature F1(Col) of the self-defined words W _A and W _B of speakers A and B, the feature list processing module 116 may calculate the mean value Mean1(Col) and standard deviation Std1(Col) of all samples in Table 1, wherein for the features {F1(1), F1(2), F1(3), F1(4), F1(5)} (for example: short-term energy value STE( ), zero crossing rate ZCR( ), starting time point Pos( ), pitch Pitch( ) and duration Duration( )), the feature list processing module 116 can calculate the mean values {Mean1(1), Mean1(2), Mean1(3), Mean1(4), Mean1(5)} and standard deviations {Std1(1), Std1(2), Std1(3), Std1(4), Std1(5)} of all samples in Table 1; and (2) the feature list processing module 116 can convert the feature F1(Col) into the feature F2(Col), So that F2(Col)=(F1(Col)-Mean1(Col))/Std1(Col). In particular, the features {F1(1),F1(2),F1(3),F1(4),F1(5)} in the temporary feature list {L_tmp} are converted into features {F2(1),F2(2),F2(3),F2(4),F2(5)} in the feature list {L}; where "Col" can represent an integer in the interval [1,5]. The feature list processing module 116 can store the conversion parameters {(Mean1(1), Std1(1)), (Mean1(2), Std1(2)), (Mean1(3), Std1(3)), (Mean1(4), Std1(4)), (Mean1(5), Std1(5))} between the features {F2(1), F2(2), F2(3), F2(4), F2(5)} and {F1(1), F1(2), F1(3), F1(4), F1(5)} in the storage device 140 for converting the temporary feature list L_tmp _U of the audio data Audio_Data _U of the audio clip Audio_Clip _U of the speaker U into the feature list L _U of the audio data Audio_Data _U in the recognition process. The feature list processing module 116 can convert the feature F _U 1(Col) in the temporary feature list L_tmp _U into the feature F _U 2(Col) in the feature list L _U so that F _U 2(Col)=(F _U 1(Col)-Mean1(Col))/Std1(Col). In particular, the features {F _U 1(1), F _U 1(2), F _U 1(3), F _U 1(4), F _U 1(5)} in the temporary feature list L_tmp _U are converted into the features {F _U 2(1), F _U 2(2), F _U 2(3), F _U 2(4), F _U 2(5)} in the feature list L _U respectively.

FIG8 illustrates a workflow of a speaker identification control scheme of the method according to an embodiment of the present invention. Steps S34 and S35 shown in FIG2 may include multiple sub-steps, such as steps S41-S43 and steps S44 and S45, respectively, wherein the one or more features may include pitch (Pitch) and duration (Duration).

In step S41, the processing circuit 110 (or the voice type classification processing module 114) can perform a filtering operation on the pitch Pitch(U) in the feature list L _U based on the database 142 based on the feature list to determine whether the audio clip Audio_Clip _U is invalid. In particular, the pitch Pitch( _U ) is filtered based on the maximum value MAX({Pitch(A)}) and the minimum value MIN({Pitch(A)}) of the pitch {Pitch(A)} in the feature list {L _A } corresponding to speaker A, and the maximum value MAX({Pitch(B)}) and the minimum value MIN({Pitch(B)}) of the pitch {Pitch(B)} in the feature list {L B } corresponding to speaker B. If MIN({Pitch(A)})<Pitch(U)<MAX({Pitch(A)}) or MIN({Pitch(B)})<Pitch(U)<MAX({Pitch(B)}), proceed to step S42; otherwise, if the audio clip Audio_Clip _U (or speaker U) is determined to be invalid, proceed to step S36.

In step S42, the processing circuit 110 (or the voice type classification processing module 114) can perform a filtering operation on the duration Duration(U) in the feature list L _U according to the database 142 based on the feature list to determine whether the audio clip Audio_Clip _U is invalid. In particular, the duration Duration( _U ) is filtered according to the maximum value MAX({Duration(A)}) and the minimum value MIN({Duration(A)}) of the duration {Duration(A)} in the feature list {L _A } corresponding to speaker A, and the maximum value MAX({Duration(B)}) and the minimum value MIN({Duration(B)}) of the duration {Duration(B)} in the feature list {L B } corresponding to speaker B. If MIN({Duration(A)})<Duration(U)<MAX({Duration(A)}) or MIN({Duration(B)})<Duration(U)<MAX({Duration(B)}), proceed to step S43; otherwise, if the audio clip Audio_Clip _U (or speaker U) is determined to be invalid, proceed to step S36.

In step S43, processing circuit 110 (or speech type classification processing module 114) can perform machine learning-based classification based on all features in feature list L _U using a predetermined classifier, such as a k-NN algorithm classifier (or "KNN classifier"), to determine whether speaker U of audio clip Audio_Clip _U is speaker A or speaker B, thereby selectively proceeding to step S44 or step S45. If speaker U is determined to be speaker A, the process proceeds to step S44; if speaker U is determined to be speaker B, the process proceeds to step S45.

In step S44, the processing circuit 110 may execute the action Action(A) corresponding to speaker A.

In step S45, the processing circuit 110 may execute the action Action(B) corresponding to speaker B.

According to some embodiments, one or more steps may be added, deleted, or modified in the workflow shown in Figure 8.

FIG. 9 illustrates the data points involved in the k-NN classifier in the speaker identification control scheme according to an embodiment of the present invention. For example, a KNN classifier operating according to the k-NN algorithm can use majority decision to perform classification when dealing with classification problems, and the relevant operations may include: (1) determining the value of k; (2) calculating the distance between each neighbor and a target data point (e.g., a new data point); (3) finding the k neighbors closest to the target data point (e.g., a new data point); and (4) checking which category/group (e.g., category A or category B) among multiple predetermined categories has the largest number of neighbors, so as to classify the target data point (e.g., new data point) into the category/group (e.g., category A or category B) with the largest number of neighbors; wherein the multiple predetermined categories, such as category A and category B, may correspond to the above-mentioned registered speakers, such as speakers A and B. The dimension of the predetermined space (e.g., the number of axes in the plurality of axes {X ₁ , X ₂ , ...}) can be equal to the number of feature types in the plurality of predetermined types of features (e.g., 5). For better understanding, FIG. 9 illustrates the plurality of axes {X ₁ , X ₂ } using two axes {X ₁ , X ₂ } as an example. Furthermore, data points belonging to class A can be represented by a feature list {L _A }, data points belonging to class B can be represented by a feature list {L _B }, and new data points can be represented by a feature list L _U . Furthermore, the number of data points CNT_Data_point _A (e.g., CNT_Data_point _A = 9) belonging to class A can be equal to the number of features _{CNT_LA} (e.g., _{CNT_LA} = 9) in the feature list { _LA }, while the number of data points CNT_Data_point _B (e.g., CNT_Data_point _B = 8) belonging to class B can be equal to the number of features CNT_LB (e.g., _{CNT_LB} = 8) in the feature list { _{LB }} . For example, for a new data point, three neighbors belong to class A and two neighbors belong to class B. In this case, the KNN classifier can classify the new data point as class A. According to some embodiments, the new data point, the data point belonging to class A, the data point belonging to class B, the number of data points CNT_Data_point _A , the number of data points CNT_Data_point _B , the number of feature lists _{CNT_LA} , the number of feature lists _{CNT_LB} , the number of neighbors belonging to class A, and/or the number of neighbors belonging to class B may vary. For example, for a new data point, there are two neighbors belonging to class A and five neighbors belonging to class B. In this case, the KNN classifier may classify the new data point as class B.

Figure 10 shows a flow chart of the method according to one embodiment of the present invention.

In step S50, the processing circuit 110 may perform the feature collection on the audio data Audio_Data (e.g., audio data {Audio_Data _A } and {Audio_Data _B }) of at least one audio clip Audio_Clip in the registration phase to generate at least one feature list L (e.g., feature list {L _A } and {L _B}) of the at least one audio clip Audio_Clip. }), to establish a feature list-based database 142 in the voice control device 100, wherein the at least one audio clip Audio_Clip may carry the at least one custom word W, and the feature list-based database 142 may include the at least one feature list L, and any feature list L in the at least one feature list L may include multiple features of an audio clip Audio_Clip corresponding to one of the at least one audio clip Audio_Clip, and the multiple features respectively belong to the multiple predetermined types of features.

In step S51 , the processing circuit 110 may perform the feature collection on the audio data Audio_Data (eg, audio data Audio_Data _U ) of another audio clip Audio_Clip (eg, audio clip Audio_Clip _U ) in the identification phase to generate another feature list L (eg, feature list L _U ) of the another audio clip Audio_Clip.

In step S52 , during the identification phase, the processing circuit 110 may perform at least one filtering operation on at least one feature in the other feature list L based on the feature list database 142 to determine whether the other audio clip Audio_Clip is invalid, thereby selectively ignoring the other audio clip Audio_Clip or performing at least one subsequent operation, wherein the at least one subsequent operation includes waking up the voice-controlled device 100.

According to some embodiments, one or more steps may be added, deleted, or modified in the workflow shown in FIG. 10 .

The above description is merely a preferred embodiment of the present invention. All equivalent changes and modifications made within the scope of the patent application of the present invention should fall within the scope of the present invention.

100: Voice Control Device 110: Processing Circuit 111: Processing Module 112: Threshold Initialization Processing Module 113: Short-Term Energy (STE) and Zero Crossing Rate (ZCR) Processing Module 114: Voice Type Classification Processing Module 115: Pitch Processing Module 116: Feature List Processing Module 120: Audio Input Device 130: Audio Data Conversion Interface Circuit 140: Storage Device 141: Programming Code 142: Feature List-Based Database

Claims (9)

A method for waking up a voice-controlled device by detecting voice features of a self-defined word, the method comprising: In a registration phase among multiple phases, collecting features of audio data of at least one audio clip; Performing feature collection on the audio data of the at least one audio segment to generate at least one feature list for the at least one audio segment, thereby establishing a feature-list-based database in the voice-controlled device, wherein the at least one audio segment carries at least one custom word, the feature-list-based database includes the at least one feature list, any one of the at least one feature list includes multiple features of an audio segment corresponding to one of the at least one audio segment, and the multiple features respectively belong to multiple predetermined types of features, wherein performing feature collection on the audio data of the at least one audio segment to generate the at least one feature list for the at least one audio segment further includes: After recording the corresponding audio segment to obtain corresponding audio data of the corresponding audio segment, analyzing first audio data of a first partial audio segment of the corresponding audio segment to determine an energy threshold and a zero-crossing rate threshold based on a plurality of first audio frames of the first audio data for further processing remaining audio data of a remaining partial audio segment of the corresponding audio segment; and Analyzing the remaining audio data of the remaining partial audio segment to calculate the energy values and zero-crossing rates of each of the plurality of second audio frames of the remaining audio data, and determining whether the speech type of any second audio frame in the plurality of second audio frames reaches the energy threshold and the zero-crossing rate of any second audio frame reaches the zero-crossing rate threshold, so as to determine the plurality of features of the corresponding audio segment based on the speech types of the plurality of second audio frames; In an identification phase among the plurality of phases, performing the feature collection on the audio data of another audio segment to generate another feature list for the another audio segment; and During the recognition phase, at least one filtering operation is performed on at least one feature in the other feature list based on the feature list database to determine whether the other audio segment is invalid, thereby selectively ignoring the other audio segment or performing at least one subsequent operation, wherein the at least one subsequent operation includes waking up the voice control device. A method as described in item 1 of the patent application, wherein the at least one audio segment comprises a plurality of audio segments, and the at least one feature list comprises respective feature lists of the plurality of audio segments, wherein any feature list in the at least one feature list represents one of the respective feature lists of the plurality of audio segments, and the corresponding audio segment represents one of the plurality of audio segments. The method as described in claim 1, wherein performing the at least one filtering operation on the at least one feature in the other feature list based on the feature list-based database to determine whether the other audio segment is invalid, thereby selectively ignoring the other audio segment or performing the at least one subsequent operation further comprises: If the other audio segment is invalid, ignoring the other audio segment; and If the other audio segment is not invalid, performing the at least one subsequent operation. The method as described in claim 1, wherein the at least one audio segment includes at least one first audio segment of a first user and at least one second audio segment of a second user; and performing feature collection on the audio data of the at least one audio segment to generate the at least one feature list for the at least one audio segment further comprises: Performing feature collection on the audio data of the at least one first audio segment to generate at least one first feature list for the at least one first audio segment, wherein each first audio segment of the at least one first audio segment contains a first custom word, the feature list-based database includes the at least one first feature list, any first feature list of the at least one first feature list includes multiple first features of a first audio segment corresponding to one of the at least one first audio segment, and the multiple first features respectively belong to the multiple predetermined types of features; and The feature collection is performed on the audio data of the at least one second audio segment to generate at least one second feature list for the at least one second audio segment, wherein each of the at least one second audio segment contains a second custom word, the feature list-based database includes the at least one second feature list, and any second feature list in the at least one second feature list includes multiple second features of a second audio segment corresponding to one of the at least one second audio segment, and the multiple second features respectively belong to the multiple predetermined types of features. The method as described in claim 1, wherein a predetermined classifier corresponding to at least one predetermined model is established in the voice control device by performing machine learning; the at least one feature in the other feature list is at least one of all features in the other feature list, wherein all features in the other feature list belong to the plurality of predetermined types of features; and the at least one subsequent operation further comprises: Based on all features in the other feature list, using the predetermined classifier to perform machine-learning-based classification to determine whether the speaker of the other audio segment is a first user or a second user, and selectively performing at least one first action corresponding to the first user or at least one second action corresponding to the second user. The method of claim 5, wherein a dimension of a predetermined space of the at least one predetermined model is equal to a feature-type count of the plurality of predetermined types of features. The method as described in claim 1, wherein performing the at least one filtering operation on the at least one feature in the other feature list based on the feature list database to determine whether the other audio segment is invalid, thereby selectively ignoring the other audio segment or performing the at least one subsequent operation further comprises: Performing the at least one filtering operation on the at least one feature in the other feature list based on the feature list database to determine whether the other audio segment is invalid, thereby selectively ignoring the other audio segment or performing the at least one subsequent operation, without requiring a network connection to any cloud database to obtain any voice data for determining the characters in the at least one custom word. The method as described in claim 1, wherein collecting features from the audio data of the at least one audio segment to generate the at least one feature list for the at least one audio segment further comprises: Dividing the corresponding audio segment into a plurality of audio segments based on the respective audio types of the plurality of second audio frames, wherein any two adjacent audio frames having the same predetermined audio type in all audio frames of the corresponding audio data belong to the same audio segment, wherein all audio frames of the corresponding audio data include the plurality of first audio frames and the plurality of second audio frames, and a leading audio segment of the plurality of audio segments includes at least the plurality of first audio frames and corresponds to a first predetermined audio type; Calculating a total duration of at least one primary audio segment from the plurality of audio segments as one of the plurality of features of the corresponding audio segment, wherein the at least one primary audio segment includes audio segments from the plurality of audio segments excluding the leading audio segment and any trailing audio segments corresponding to the first predetermined speech type; and Calculating at least one segment-level parameter for each audio segment from the plurality of audio segments corresponding to a second predetermined speech type, and determining at least one parameter of the corresponding audio segment based on the at least one segment-level parameter as at least one other feature of the plurality of features of the corresponding audio segment. A processing circuit for waking up a voice-controlled device by detecting voice features of self-defined words. The processing circuit comprises: A plurality of processing modules for operating the processing circuit, wherein the plurality of processing modules comprise: A feature list processing module for performing feature-list-related processing; and At least one other processing module for performing feature collection; Wherein: In a registration phase among a plurality of phases, the processing circuit performs the feature collection on audio data of at least one audio clip. Performing feature collection on the audio data of the at least one audio segment to generate at least one feature list for the at least one audio segment, thereby establishing a feature-list-based database in the voice-controlled device, wherein the at least one audio segment carries at least one custom word, the feature-list-based database includes the at least one feature list, any one of the at least one feature list includes multiple features of an audio segment corresponding to one of the at least one audio segment, and the multiple features respectively belong to multiple predetermined types of features, wherein performing feature collection on the audio data of the at least one audio segment to generate the at least one feature list for the at least one audio segment further includes: After recording the corresponding audio segment to obtain corresponding audio data of the corresponding audio segment, analyzing first audio data of a first partial audio segment of the corresponding audio segment to determine an energy threshold and a zero-crossing rate threshold based on a plurality of first audio frames of the first audio data for further processing remaining audio data of a remaining partial audio segment of the corresponding audio segment; and Analyzing the remaining audio data of the remaining partial audio segment to calculate the energy values and zero-crossing rates of each of the plurality of second audio frames of the remaining audio data, and determining whether the speech type of any second audio frame in the plurality of second audio frames reaches the energy threshold and the zero-crossing rate of any second audio frame reaches the zero-crossing rate threshold, so as to determine the plurality of features of the corresponding audio segment based on the speech types of the plurality of second audio frames; In an identification phase among the plurality of phases, the processing circuit performs the feature collection on the audio data of another audio segment to generate another feature list for the another audio segment; and During the recognition phase, the processing circuit performs at least one filtering operation on at least one feature in the other feature list based on the feature list database to determine whether the other audio segment is invalid, thereby selectively ignoring the other audio segment or performing at least one subsequent operation, wherein the at least one subsequent operation includes waking up the voice control device.